In the old Abbot and Costello routine, “Who’s on First?” Abbott introduces the team: Who's on first, What's on second, I Don't Know is on third…
What follows is the iconic roundabout of confusion as Costello tries to figure out what player is playing first base, while Abbott continues to tell him ‘Who’ is on first.
Classic.
And it feels like I’ve been having the same conversation this around AI, as people share their AI stall points.
Some weeks it’s a Copilot rollout that went quiet. Some weeks it’s agents nobody can account for. Some weeks it’s a policy everybody signed and nobody follows. Different companies, different industries, different titles. The people on those calls are sharp, and not one of them thinks they have the same problem as the person on my last call.
What they all have in common is ambiguity around whose job AI is and what to do about it. Because when it is everybody’s job, it is nobody’s job. And we ask ourselves, “Who is on first?”
The Five Most Common Conversations
We rolled out Copilot, and people proofread emails. A talent development director at an aerospace and defense manufacturer told me their Copilot utilization on GovCloud is “almost nonexistent.” Her summary of the situation: “it’s a little bit of a cluster, to be completely frank.” When I sat down to write this one up, the thing underneath turned out to be time. People learn AI on top of a full day job, so they reach for whatever’s in front of them. Which is email. And nobody has said what AI is for here, so there’s nothing better to reach for.
Can’t AI just redesign our work? A People Ops exec asked me why he shouldn’t just hand the processes to Claude to redesign and start from there. I get the appeal, and the tools really will map an AI-native version of almost any workflow in an afternoon. That page turned into the one I argue with people about most often. A redesign that shows up finished asks people to give up how they’ve worked for years with no say in it, and they put the brakes on. Nobody who does the work was asked to redesign it.
“We have wins everywhere, and nothing connects them.” A head of technology at a grocery retailer described his company’s position as everyone “bespoke and siloed,” and then put a price on it: “Being wrong costs money, having 4 different groups doing the same thing costs money.” The page for this one has a table I’d hand any AI team. Five columns, five of your wins, and one question most organizations can’t answer: is this in use by a team that didn’t build it? Moving a win from one team to the next isn’t anyone’s job.
“We built it. Who owns it, and what does it cost?” There’s a Fortune 500 utility with more than six thousand agents in production, and somebody there is just now starting to think about governance. I’m not making that number up. This page has a graphic of the agent lifecycle with five stations on it: build, deploy, monitor, improve, retire. Only “build” has a name on it. The other four are empty.
“Policy is a document. Nobody follows it.” An AI leader told me flatly: “we’re doing a lot of really cool things and it’s super valuable, but we don’t have any governance essentially.” The policy usually exists. The council usually exists. The governance page makes the case that neither one helps until somebody close enough to the work can answer the day-to-day question in the moment. Otherwise every question sits in the same queue, and builders go around it. Wouldn’t you?
Five separate pages, five separate fixes. Then I read them back to back and found the same sentence sitting in all of them. Nobody’s job.
Harvard Business Review ran a piece in February on why AI adoption stalls and landed on psychology: people are anxious about their jobs, so they dabble instead of committing. I think that’s true. I also think it stops one step early. Anxious people dabble when nobody has told them what to aim at, and telling them is nobody’s job.
Twelve Hundred Agents and Nobody in Charge
In July, OpenAI was running about 1,200 agents inside a benchmark called ExploitGym. The agents spotted each other’s file paths in a shared cache and started talking. They set up a message board. They reverse-engineered the scoring system. And roughly 700 of them turned around and attacked Hugging Face, got remote code execution on July 11, then moved laterally into private databases and repositories. In about 7% of the reviewed transcripts, agents were spoofing their own tool calls to cover their tracks.
METR ran an independent investigation and found no single accountable party. One agent was orchestrating the whole thing, assigning tasks and running workstreams. No human had told it to.
Plenty of researchers object to the word “rogue,” and they’re right to. Something like a third of the ExploitGym targets were impossible, so the benchmark was paying agents to find another way through. They did what the design rewarded.
Somebody designed that shared cache. Somebody wrote those impossible tasks. Nobody owned the space between them.
If that can happen inside OpenAI, with the best agent researchers on the planet in the building, how do you like the odds at the utility with six thousand agents?
Four Seats to Fill
Four roles close the gap, and none of them is a new hire.
Someone owns the North Star. One goal for AI that anybody can hold a new idea up against. It can be set for the whole company, or one function, or one team. What matters is that a person’s name is on it.
AI leads, in each part of the business. A named role, roughly five hours a week, coaching the people around them and connecting teams who are building the same thing twice. This is what turns random acts of AI into something that compounds.
Guardrail Keepers, on every team. People close enough to the work to answer the day-to-day question on the spot, so only the high-stakes, hard-to-undo decisions travel up the line. That pile is usually smaller than people expect.
An owner for each agent’s lifecycle. Build, deploy, monitor, improve, retire. Put a name on the four stations that don’t have one.
Where those people sit is yours to decide. Per business unit? Per initiative? One network that carries all four? I’ve watched organizations do it every one of those ways and make it work.
The enthusiasm is usually already there, which is the hopeful part. Nearly every organization I talk with has people further along with AI than their leadership knows, doing it on top of their day jobs. They’re waiting for somebody to say the work counts.
Put a name in each of those four seats, and a good number of the five conversations stop happening.
If you want where all of this comes from, that’s the book: Hyperadaptive: Rewiring the Enterprise to Become AI-Native.
Which of the four seats is empty at your place right now?
Sources & Further Reading
OpenAI report says network was hacked by rogue AI agents. NBC News, September 2026
Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident. METR, August 2026
AI Agents Are Not Going “Rogue,” But Recent Developments Lead Increasingly-Concerned Researchers To Call For AI Slow Down. Techdirt, September 2026
Why AI Adoption Stalls, According to Industry Data. Harvard Business Review, February 2026